At the Toyota Research Institute (TRI), researchers are showcasing a new method that enables robots to learn new skills overnight. This breakthrough in robot learning is achieved by combining traditional learning techniques with diffusion models. TRI has successfully trained robots on 60 skills using this method, which is based on the premise that existing models alone cannot solve the problem of teaching robots specific tasks that require physical interaction.
The advantage of this method is that it allows the programming of skills that can function in diverse settings, overcoming the challenge of robots struggling to perform in unstructured environments. This is particularly valuable in scenarios such as homes, where robotic systems need to adapt to changing conditions and navigate through obstacles like furniture or messes left by humans.
Traditionally, roboticists have relied on programming robots to anticipate and manage various edge cases and deviations. However, the new approach aims to create “general-purpose” systems that can adapt and learn new tasks instead of being limited to performing a single task repeatedly.
To teach robots new skills, TRI initially uses teleoperation, where a person demonstrates the desired task while the robot imitates the actions. This process involves several hours of repetition and can be enhanced by transmitting force feedback between the robot and the operator, enabling a better understanding of the robot’s interaction with the world.
The system leverages data from various sensory inputs, such as sight and force feedback, to build a comprehensive understanding of the task. For example, tactile sensing is crucial for tasks like flipping pancakes and rolling dough, as it enhances the success rate significantly.
Once the initial training phase is complete, the neural networks of the robots continue training overnight, refining the skills learned during teleoperation. The goal is that the robot will have fully mastered the skill by the time researchers return to the lab the next day.
This advancement in robot learning is a significant step toward creating robots that can learn and adapt to new tasks in real-world environments. By combining traditional learning techniques with diffusion models, TRI is pushing the boundaries of what robots can achieve, eliminating the need for extensive training and enabling a more rapid acquisition of skills.
Sources:
– TechCrunch Disrupt
– Toyota Research Institute